Issue with Sigmoidnet input function

This is the code I am using for generating the predicted output ('yo' as mentioned in the code) from given input ('u' as mentioned in the code) and output ('y' as mentioned in the code). Though I am getting 100% data fitting during training, the trained parametric sigmoidnet model ('sys' as mentioned in the code) is predicting completely different output from the same data used for training. Can soemone please help to find out bugs in the code if there are any? Furthermore, is there any alternative method to generate predicted output from given input using sigmoidnet function?

The code is written below:

data=iddata(y,u);

sys = nlarx(data,[2 2 1],sigmoidnet('NumberOfUnits',3))

% u is the input to the model

% y is true output of the model

% x is the regressor value vector, na=nb=2 and nk=1, so number of

% regressor=4

% yo is the predicted output

P = sys.Nonlinearity.Parameters.LinearSubspace;

r = sys.Nonlinearity.Parameters.RegressorMean;

L = sys.Nonlinearity.Parameters.LinearCoef;

Q = sys.Nonlinearity.Parameters.LinearSubspace;

b = sys.Nonlinearity.Parameters.Dilation;

c = sys.Nonlinearity.Parameters.Translation;

a = sys.Nonlinearity.Parameters.OutputCoef;

d = sys.Nonlinearity.Parameters.OutputOffset;

yo = [];

for i=1:size(u,1)

x= zeros(1,4);

if i==1

x(1)=0;

x(2)=0;

x(3)=0;

x(4)=0;

elseif i==2

x(1)=data.y(i-1);

x(2)=0;

x(3)=data.u(i-1);

x(4)=0;

else

x(1)=data.y(i-1);

x(2)=data.y(i-2);

x(3)=data.u(i-1);

x(4)=data.u(i-2);

end

y_element = (x-r)*P*L + a(1)*1/(exp(-((x-r)*Q*b(:,1) + c(1)) +1)) ...

+ a(2)*1/(exp(-((x-r)*Q*b(:,2) + c(2)) +1)) ...

+ a(3)*1/(exp(-((x-r)*Q*b(:,3) + c(3)) +1)) ...

+ d;

disp(y_element)

yo = [yo; y_element];

end

% Final predicted output for display

out_predict = yo;